Neural network to identify hand-written digits, works well with the MNIST train dataset (97% accuracy) however doesn't perform as expected when input is given through the pygame mini-program, mostly because user's handwriting deviates from the dataset that the model is accustomed to.

Canvas is of size 140x140 but can be changed to meet preferences, make sure to change array_temp1 and te_pixels (twenty-eight pixels) accordingly. Thanks!